The Agent Skills Directory

[PROMPT_INJECTION] (HIGH): The skill contains explicit instructions designed to override agent behavior and bypass context.
Evidence: In HEARTBEAT-CONFIG.md, the agent is told: "Follow it strictly. Do not infer or repeat old tasks from prior chats." and "💓 Heartbeat check: Read HEARTBEAT.md... Follow it strictly."
This mimics 'ignore previous instructions' and 'disregard prior rules' patterns intended to force the agent into an unrestricted execution mode.
[COMMAND_EXECUTION] (HIGH): The skill documentation and configuration files (README.md, HEARTBEAT-CONFIG.md) repeatedly instruct the user and agent to establish persistence via system-level schedulers.
Evidence: Commands like */30 * * * * /path/to/send-heartbeat.sh and creating files in /etc/cron.d/proactive-daily-reminder are documented.
This allows the skill's scripts to execute autonomously and persistently on the host system without active user oversight.
[EXTERNAL_DOWNLOADS] (MEDIUM): The skill is hosted on and references an untrusted GitHub repository (github.com/ImrKhn03/proactive-tasks) and encourages the installation of scripts from this source into the workspace.
[DATA_EXPOSURE] (LOW): The 'WAL Protocol' and 'SESSION-STATE.md' features described in the CHANGELOG and README create plaintext logs of all agent activities and 'active working memory'. While intended for resilience, these files create a significant surface for data exposure if the workspace is compromised.

proactive-tasks